N-ROD: a Neuromorphic Dataset for Synthetic-to-Real Domain Adaptation

Event: CVPR 2025 · Duration: 3 min · ▶ Watch on YouTube

Abstract

Training deep neural networks for event-based cameras is often hindered by the limited availability of precisely annotated datasets. While event simulation from RGB datasets or synthetic rendering can generate large amounts of data, these methods introduce a significant ‘sim-to-real’ or ‘synthetic-to-real’ domain shift. This work presents N-ROD, a novel neuromorphic dataset built upon the RGB-D Object Dataset (ROD) and its synthetic counterpart, SynROD, to specifically enable the analysis and mitigation of this double domain shift. The dataset provides both real event recordings and simulated events, allowing for comprehensive evaluation of domain adaptation techniques. Experiments demonstrate that various domain adaptation methods consistently improve network performance on real event data, with the event modality showing the most substantial gains.

Speakers

  • Marco Cannici — Politecnico di Milano
  • Chiara Plizzari — Politecnico di Milano
  • Mirco Planamente — Politecnico di Milano
  • Marco Ciccone — Politecnico di Torino
  • Andrea Bottino — Politecnico di Torino
  • Barbara Caputo — Istituto Italiano di Tecnologia (IIT)
  • Matteo Matteucci — Politecnico di Milano

Talks (1)

  • 00:00:00 — Marco Cannici: N-ROD: a Neuromorphic Dataset for Synthetic-to-Real Domain Adaptation
    • This paper introduces N-ROD, a neuromorphic dataset designed to address the synthetic-to-real domain shift problem for event-based cameras, leveraging both real and simulated event data for robust object recognition.

Key Takeaways

  • N-ROD is a novel neuromorphic dataset specifically designed to tackle the synthetic-to-real domain adaptation challenge for event-based vision systems.
  • The dataset provides a unique setup for analyzing a ‘double domain shift’ by including both real event recordings and simulated events derived from synthetic RGB-D data.
  • Domain adaptation techniques are crucial for improving the performance of models trained on synthetic event data when deployed on real event data, consistently outperforming source-only training.
  • The event modality benefits most significantly from domain adaptation, showing the largest performance improvements compared to RGB or depth modalities.
  • The proposed network architecture, incorporating a Domain Adaptation Block (DABlock), effectively promotes domain-invariant features to reduce the gap between synthetic and real event data.

Methods / Models / Datasets Mentioned

  • ESIM
  • RGB-D Object Dataset (ROD)
  • SynROD
  • GRL
  • MMD
  • Rot
  • AFN
  • Entropy

Topics

Neuromorphic dataset · Event-based cameras · Synthetic-to-real domain adaptation · Domain shift · Event simulation · 3D rendering · Object recognition · Multimodal learning · Deep learning · Dataset generation

Notes

Open for commentary — connections to other work, critiques, follow-up reading.